Feynman's Path Integral for Understanding Diffusion Models
Core Concepts
Diffusion models are redefined using Feynman's path integral, bridging quantum physics and generative models.
Abstract
The content explores the application of Feynman's path integral to diffusion models. It delves into the formulation, derivation of equations, and the role of noise in sampling processes. The analysis includes experiments on Gaussian data and synthetic distributions, showcasing the impact of noise on model performance.
Introduction to diffusion models and their applications.
Formulation using Feynman's path integral for comprehensive descriptions.
Derivation of reverse-time SDEs and loss functions for training.
Perturbative evaluation of negative log-likelihood based on WKB expansion.
Experiments on 1D Gaussian data and 2D synthetic data to analyze noise effects.
Limitations and future directions discussed.
Understanding Diffusion Models by Feynman's Path Integral
Stats
"Score-based diffusion models have proven effective in image generation."
"Stochastic sampling schemes require more function evaluations than deterministic ones."
"Improvements in metrics like Fr´echet Inception Distance reported with stochastic processes."
Quotes
"The analogy between h and ℏ naturally leads us to explore a perturbative expansion in terms of h."
"Depending on the choice of ϵ, the NLL could be an increasing function of h as well."